Research On Automatic Tomato Picking Method Based On Deep Learning And Binocular Vision

Tomato picking is a necessary process for tomato processing.The quality of tomato picking will directly affect the subsequent processing of tomatoes.The yield of tomatoes is large,and the picking window period of mature tomatoes is short,so it is necessary to pick tomatoes quickly and intensively.At present,tomato picking is mainly done manually,but tomato picking is seasonal,and the picking work consumes a lot of human resources,and manual picking is subjective,which is not conducive to product quality control.For automated harvesting,tomatoes grow in dense foliage,and tomatoes appear in different sizes,colors and shapes during their own growth cycle,therefore,it is very difficult to quickly and accurately identify and locate ripe tomato targets in scenes with complex backgrounds.The traditional machine vision algorithm is greatly affected by illumination,occlusion and background,and has poor robustness,which cannot fulfill the actual production requirements of tomato picking.In deep learning,the tomato target is a small target,the image features are not obvious,and the existing deep learning models are prone to missing identification.Therefore,in view of the problems in the tomato picking process,the research on the recognition method of ripe tomato fruits based on deep learning is carried out,and further research is carried out.Research on the spatial positioning method of the identified ripe tomatoes.Thesis analyzes the SSD model architecture,adding a shallower Conv3＿3 feature layer;introducing feature pyramid feature fusion and continuous feature layer upsampling structure;at the same time,the size of the a priori box is changed to make it more suitable for the size of the actual targets.In the case of reducing a certain detection speed,the SSD model with improved structure has a 22.1% increase in detection accuracy,which significantly improves the detection rate of tomato by the original model.On the basis of the improvement of the model structure,the rectangular a priori box of the model is replaced with a circular a priori box,and the calculation method of the intersection ratio and loss function is improved.The detection accuracy of the improved model is further improved by 1.1%.The accuracy of the position of the tomato is increased by 21.9%,and the improved model has a better regression effect on the edge of the tomato,providing more accurate image coordinates for the spatial positioning of the tomato.Based on binocular vision,thesis conducts experiments on binocular targeting,stereo matching,and RGB image and depth map registration.On the basis of RGB-D image,a method for solving the transformation relationship of spatial coordinate system using augmented matrix is proposed,Experiments show that the maximum absolute error of this method is 9.775 mm and the maximum relative error is 1.298% in tomato space positioning.The tomato automatic picking integrated system is established,the tomato target detection system and the tomato target positioning system are integrated,the tomato image coordinates of the tomato target detection system are transmitted to the tomato target positioning system through the local file,the system is debugged,and the test is completed.